List of Papers Browse by Subject Areas Author List
Abstract
Cardiac magnetic resonance (CMR) is a clinical reference standard for assessing cardiac morphology and function. Whole-heart assessment relies on sparse multi-plane cine acquisitions from complementary short-axis (SAX) and 2/3/4-chamber long-axis (LAX) views across the cardiac cycle. Although large-scale cohorts such as UK Biobank provide tens of thousands of such studies, most scans remain unlabeled, and learning compact whole-heart representations from the resulting high-dimensional 3D+t data is still challenging under practical memory constraints. To address this, we introduce a geometry-aware self-supervised learning framework for whole-heart representation learning from sparse multi-plane cine CMR. First, we inject patient-space coordinates into spatio-temporal tubelet tokens, which preserves explicit spatial identity and enables shared per-plane encoding. Next, we explicitly model cross-plane structure by aggregating multi-plane features into a latent 3D volume via voxel-query cross-attention. To ensure geometric consistency while keeping supervision efficient, we reconstruct masked tubelets by coordinate-constrained trilinear reprojection from the latent volume. Finally, we adopt hierarchical masking that combines intra-plane tubelet masking with inter-plane plane dropping, turning reconstruction into cross-plane completion and strengthening global 3D reasoning. Experiments show that coherent 3D anatomy can be decoded from the learned latent volume using a lightweight linear probe, and that masked-plane reconstruction recovers missing views from remaining planes. Moreover, our method improves downstream multi-plane segmentation over state-of-the-art methods, while maintaining favorable efficiency under the same input setting. Code is available at this \href{https://github.com/Liwei-Hu1/Geometry_Aware_CMR_Pretraining.git}{URL}.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5507_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Liwei-Hu1/Geometry_Aware_CMR_Pretraining.git
Link to the Dataset(s)
N/A
BibTex
@InProceedings{HuLiw_GeometryAware_MICCAI2026,
author = { Hu, Liwei AND Wang, Zi AND Wu, Yinzhe AND Zhang, Zhenxuan AND Wang, Anbang AND Yuan, Ruicheng AND Wang, Fanwen AND Yap, Choon Hwai AND Yang, Guang},
title = { { Geometry-Aware Self-supervised Learning for Whole-Heart 3D+t Representation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16889},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes a geometry-aware SSL framework designed to learn 3D+t representations from sparse, multi-plane cine CMR. The authors introduce patient-space information to encode sparse and unordered multi-plane cine CMR. The authors also adopted a volume modelling module to enforce geometry-consistency. The authors introduce hierarchical masking, intra-plane tubelet masking and inter-plane plane dropping, to force the model to perform cross-plane anatomical reasoning.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
-
The integration of physical patient-space coordinates (multi-modality information) into the transformer architecture is a principled approach to handling sparse, intersecting medical imaging planes.
-
The plane-to-volume and volume-to-plane projection enforces a strong geometric prior. The plane-to-volume module also seems a good translational adaptation of VoxFormer.
-
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.Motivation and Efficiency: The authors state that the primary motivation is to utilise the unlabeled data in datasets via SSL. However, it remains unclear whether the ultimate goal is to propose a “Cardiac Foundation Model” for various downstream tasks or simply to achieve label-efficiency for standard segmentation. If the goal is strictly segmentation performance, the architectural complexity, which involves patient-space coordinate injection, voxel-query cross-attention, and trilinear reprojection. It appears high for the incremental gains shown in Table 1.To claim the “favourable efficiency,” the authors should report total training FLOPs or total parameter sizes, comparing to compared methods. Specifically, it is necessary to demonstrate that SSL pre-training +”lighthead” fine-tuning offers a genuine advantage over training a strong, established baseline like nnU-Net from scratch.
2.Experimental Validation and Representation Analysis: In the Representation Analysis subsection, the authors show that the model performs cross-plane completion rather than simple 2D inpainting. While this proves the model has learned a spatial prior, there is a risk that the model is simply generating an “average” heart shape that is anatomically plausible but not patient-specific. The potential clinical application of this “cross-view inpainting” is not sufficiently discussed. The authors should demonstrate how this capability benefits specific clinical workflows, such as maintaining diagnostic accuracy in the event of missing clinical views or improving the robustness of automated pipelines faced with sparse acquisitions.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
- in Eq (4), e_{vis} refers to shared encoder right? the notation of it wasn’t explained.
- in Fig 1, as you described the input to recon decoder is c_{i,j} and V, which means the inputs in Phase I to the decoder should ve illustrated as (darker blue + lighter blue), respectively? You illustrated as (colour + dark grey = V + masked tokens)? Also, dark grey appears in the previous column too.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The preliminary recommendation of Weak Accept is justified by the following major factors: The proposed framework is methodologically innovative and well described in the paper. However, as mentioned in the major weakness, I believe the authors need to justify the motivation and application of the proposed work.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This paper presents a geometry-aware SSL framework for sparse multi-plane cardiac MRI. It injects DICOM physical coordinates and performs plane-to-volume cross-view aggregation and reprojection in a latent space to address alignment and computational limits in whole-heart 3D+t representation learning.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.It replaces implicit view embeddings with patient-space geometric coordinates. 2.It keeps the reconstruction objective in the tubelet space rather than the voxel space. 3.Compared with WholeNet, the method improves performance on all five anatomical structures and overall reaches or approaches nnUNet.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.Please report statistical significance tests. 2.Please provide ablations for physical coordinate injection, explicit plane-to-volume modeling, and hierarchical masking. 3.Please report comparisons with plain MAE baselines. 4.Please report comparisons with CMR-specific foundation models such as CineMA[1].
[1] Cinema: a foundation model for cine cardiac mri
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
please refer to the weaknesses
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
the rebuttal addresses most of my main concerns
Review #3
- Please describe the contribution of the paper
This paper proposes a geometry-aware self-supervised framework for learning whole-heart 3D+t representations from sparse multi-plane cine CMR. The method injects patient-space coordinates into tubelet tokens, aggregates multi-plane features into a latent 3D volume via voxel-query cross-attention, and reconstructs masked regions through coordinate-constrained trilinear sampling. A hierarchical masking strategy encourages cross-plane reasoning, and the learned representation is evaluated via reconstruction, probing, and downstream segmentation.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The motivation is well grounded, and leveraging DICOM geometric information to enable cross-plane reasoning is natural and intuitive. The overall framework is conceptually clean, and the experimental evaluation is relatively comprehensive.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The plane-to-volume aggregation is still largely implemented via generic cross-attention, although the spatial relationship between voxel queries and tubelet tokens is explicitly available through patient-space coordinates. It may be beneficial to explore more geometry-constrained aggregation strategies, such as distance-aware attention or coordinate-based deterministic lifting, which could better preserve spatial consistency in the latent volume and make the subsequent trilinear interpolation assumption more justified.
2.The method appears to assume that anatomical structures are roughly aligned across subjects in patient-space coordinates. It would be helpful to discuss whether any preprocessing is required to ensure that corresponding chambers fall into similar regions of the latent grid, especially when handling heterogeneous acquisition protocols.
3.Since the method models spatiotemporal information, it would be helpful to include additional experiments evaluating temporal consistency, cross-phase reconstruction, or missing-frame completion to better demonstrate the learned temporal dynamics.
4.Some details in Fig. 1 are unclear. For example, the meaning of the colored border around “Geometry-injected visible tokens” is not explained, and it is ambiguous whether masked tokens are also fed into the Volume Modelling Module or only used during reconstruction. Clarifying these visual elements would improve readability and make the pipeline easier to follow.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Although some details of the method still need further clarification, this paper proposes a well-motivated and practical geometry-aware self-supervised framework and provides relatively reliable experimental verification, making an overall significant contribution.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank the reviewers for recognizing that our motivation is “well grounded”(R3), and that our geometry-aware design uses patient-space/DICOM coordinates to replace implicit view embeddings(R2), providing a principled geometric prior for cross-plane reasoning under sparse intersecting cine planes(R1,R3).
Q1: Clarifying the primary motivation(R1) A1: The motivation is to learn an effective geometry-aware whole-heart 3D+t representation for downstream tasks such as segmentation, motion tracking, image enhancement. It is not for segmentation alone, although we use it as a proxy test. We don’t consider it a foundation model either, since it is very specific to geometry/motion, and uses limited samples.
Q2: Regarding “favourable efficiency”(R1) A2: The efficiency advantage is relative to global fusion over all multi-plane tokens. Sec.2.2 shows that such global fusion has O(S^2n^2) attention complexity, whereas our shared per-plane encoding reduces this to O(Sn^2), with cross-plane interaction deferred to a compact latent volume. Thus, the explicit geometry prior is used to improve fusion efficiency, not to add complexity.
Q3: Is the representation analysis just generating an “average heart”(R1)? A3: It’s not because it is difficult for an average heart encoding to be decoded accurately into very different images of many individuals. We further observed substantial diversity in the 3D encoding space across subjects, suggesting that the network learns subject-conditioned rather than average-heart representations.
Q4: Clinical relevance of the representation analysis(R1) A4: Relevant applications include correcting MRI breath-hold slice-misalignment artefacts, improving 3D shape/motion quantification by using LAX/SAX input, and sparse-view 3D reconstruction to reduce the number of acquired slices and scanning time.
Q5: Regarding ablation studies(R2) A5: Our components jointly address one central problem: learning coherent 3D+t representations from sparse multi-plane cine CMR. DICOM coordinates provide the shared spatial frame, plane-to-volume modelling forms the latent 3D representation, and hierarchical masking forces cross-plane completion rather than 2D inpainting shortcuts. Existing results support them: 1)Fig.2(a) shows coherent 3D anatomy decoded from the latent volume, 2)Fig.2(b) shows missing-plane reconstruction from remaining views, and 3)Table 1 shows consistent gains over WholeNet under the same input setting. Since these components define the proposed geometry-aware formulation, removing any of them would change the intended representation setting rather than provide a meaningful ablation.
Q6: Requests for comparison with MAE-style methods and CineMA(R2) A6: We selected WholeNet as the closest baseline because it also uses multi-plane SAX/LAX inputs and learns a 3D+t representation. In contrast, plain MAE-style methods are designed for single-image/video inputs and do not model multi-plane inputs or 3D geometry. CineMA is an impressive CMR foundation model for 2D+t cine images, but lacks an explicit 3D representation. Table 1 shows that our explicit 3D representation improves performance while avoiding computationally expensive global attention across planes.
Q7: Additional methodological clarifications(R2,R3) A7: Temporal consistency evaluation is useful but remains future work. For distance-aware constraints, our voxel-query cross-attention learns geometry-conditioned correspondences from patient-space coordinates, while hard distance matching is suboptimal for sparse SAX/LAX cine CMR due to partial-volume effects, slice thickness, and view-dependent boundary offsets. For cross-subject alignment, our method does not assume identical chamber locations, but uses each subject’s DICOM metadata as input. Wilcoxon signed-rank tests on the reported Dice scores in Tables 1–2 show statistical significance and support our conclusions.
Minor modifications: We will clarify Eq. (4) notation and improve Fig.1 readability.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
A further rebuttal is required for the reviewers to reach a consensus.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Good work.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper presents a insightful approach by incorporating explicit geometric priors into multi-plane CMR self-supervised learning. Its strength lies in addressing the alignment of orthogonal planes using physical coordinates. The main concerns include motivation clarification, results explanation, performance comparison, and ablation study. The author’s response address most of them and all reviewers agree to accept the paper.